pi-omplike-advisor

pi-omplike-advisor

A persistent advisor extension for pi: a second model that reviews the main agent's work each turn and injects concise advice inline. Port of oh-my-pi's advisor onto upstream pi's public extension surface.

What it does

The advisor is a long-lived, read-only Agent with its own model and read-only tools (read/grep/find) plus one advise tool. It is fed the primary agent's transcript one turn-delta at a time and may inject short advice back into the conversation. It is not an executor: it cannot edit files, run commands, or change session state.

All severities enter one pending-advice queue; turn boundaries and advisor-review completion are the only places that flush it. Delivery policy then differs by severity:

• nit — tagged as raised about an earlier step. If observed while the assistant is responding, it waits for that turn's boundary (Pi would not insert a steer before then anyway). A non-terminal turn flushes it before the next step; a terminal turn sends it through final-review reconfirmation, so obsolete lagging advice is dropped and surviving advice lands with a request for a fresh, self-contained final answer. If the advisor cannot reconfirm in time, only concerns/blockers ship best-effort; unconfirmed nits remain queued. A successful late review reconciles them, otherwise the next turn boundary applies the normal low-stakes nit policy.
• concern / blocker — always held on first emission, never steered immediately. Because review is asynchronous (seconds), high-severity advice is usually stale by the time it could land, so it is held and re-confirmed by the next review (the advisor re-raises survivors and stays silent on resolved ones).

Context management (self-compaction)

The advisor accumulates one turn-delta per primary turn in its own context. It self-compacts so long sessions keep getting reviewed instead of silently failing once its context fills:

• Proactive — before each review, if the advisor's own context has crossed ADVISOR_COMPACT_AT% of its window, it clears its message history and replays the incoming turn fresh.
• Reactive — if a review still overflows mid-stream (stopReason "length"), it clears its history and replays that batch once into a fresh context. If a fresh replay still overflows, the single batch genuinely doesn't fit and the review is dropped as failed (no infinite retry).

Held concern/blocker notes are not lost across a self-compaction: they live outside the agent transcript and ride the next review as the reconfirm preamble. This is independent of the primary's own compaction, which still triggers a full reset() of the advisor.

While a high-severity note is held — or whenever a turn is about to idle — the primary's next step is stalled (a catch-up block) so the advisor can catch up. The wait backs off 15s → 30s → 60s … capped at 120s, is Escape-abortable, and shows a notice. Nothing here is ever a hard interrupt; abort() is never called.

Installation

Add the package to your pi settings (~/.pi/agent/settings.json):

{
  "packages": [
    "packages/pi-omplike-advisor"
  ]
}

(Or install from npm / a git checkout the same way you install other pi packages.)

Usage

• /advisor or /advisor status — show whether the advisor is on and which model it is using.
• /advisor on — enable the advisor (persisted; it is on by default unless explicitly turned off).
• /advisor off — disable the advisor (persisted).

Configuration

Advisor model

The advisor model defaults to openrouter/z-ai/glm-5.2. Override it by adding an advisor entry to modes.json (project .pi/modes.json or global ~/.pi/agent/modes.json):

{
  "modes": {
    "advisor": {
      "provider": "openrouter",
      "modelId": "z-ai/glm-5.2",
      "thinkingLevel": "low"
    }
  }
}

If the configured model can't be resolved, the advisor falls back to the current session model.

System prompt

The advisor ships with a built-in default system prompt (a string constant in extensions/advisor.ts). To override it, drop your own ~/.pi/agent/system-prompts/advisor.md — if present, it replaces the bundled default wholesale.

Project guidance (WATCHDOG.md)

If a WATCHDOG.md file exists in the working directory, its contents are appended to the advisor's system prompt as advisor-only guidance (review priorities, project traps, etc.). This lets you tune what the advisor watches for without touching the main agent's prompt.

Environment variables

• ADVISOR_DEBUG=1 — verbose debug logging.
• ADVISOR_COMPACT_AT=80 — % of the advisor's context window at which it proactively self-compacts (clamped to 50–95; default 80).
• ADVISOR_NO_REVIEW=1 — skip live model review (keeps the deterministic /advisor test delivery path). Used by the test harness.

Development

# fast, offline unit + loader + render tests
node packages/pi-omplike-advisor/extensions/advisor.test.mjs

# also run the live pi E2E harness (needs anthropic auth + network)
ADVISOR_E2E=1 node packages/pi-omplike-advisor/extensions/advisor.test.mjs

The harness locates pi via readlink -f $(command -v pi); if your pi is a wrapper script instead of a symlink, point PI_DIST at the install's dist/ directory (e.g. PI_DIST=~/src/pi-mono/packages/coding-agent/dist).